parser+source: improvements and cleanups

[feds01]

Changes

This PR is an aggregation of various improvements and cleanups to the parser and constants API in the compiler.

Improvements to constant storage:

• Move out methods from CONSTANT_MAP and just place them onto each Interned<ConstKind> - this improves code readability in quite a few places.
• Don't use a HashMap for storing floats/ints as they don't need to be looked up in that manner.
• use fxhash to look up identifiers and strings instead of the default dashmap hasher.

Improvements to the parser:

• Use a more expressive ExpectedItem type to represent what the parser might be expecting when parsing items, which gives it more flexibility to specify expected items other than tokens. Additionally, it saves on quite a bit of memory around the place: the size of ParseResult<T> went down from a minimum bound of 70 bytes to 40 bytes.
• Use a perfect hash function to lookup if an identifier is a keyword, there should be a 5% improvement in speed to lexing (there is more that could be done here, maybe use gprerf).
• Collect tokenisation and generation timings, and report them as other stages do.
• remove is_compound_expr flag on AstGen.
• Use ThinVec for diagnostics across compiler, which should reduce several items that used to store Vecs inline (which seemed wasteful). It also optimises for the common case of no diagnostics which effectively means that it will be a std::ptr::nullptr() and only takes up 8bytes instead of 24bytes as a Vec would.
• Share diagnostics among all AstGens which again reduces the size of an AstGen from 112 bytes to 64 bytes.
• parser: fix bug where errors were being entirely ignored - due to not consuming errors when they should be.
• parser: fix span of ParseWarning::RedundantParenthesees to be more accurate.

Todo

• ast: use TyParams and Params to represent parameters in various places #931 needs to be merged before review.

[feds01]

This should be removed